Fix MXFP8-training related issue

[WanZzzzzz]

Description

Fix 2 bugs in mxfp8 training.

(1) The following error happened after validation since colwise data is cleared by update_usage() during validation.

0: [rank0]:   File "/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/tensor/_internal/mxfp8_tensor_base.py", line 192, in update_usage
0: [rank0]:     raise RuntimeError(
0: [rank0]: RuntimeError: Requested column-wise usage, but MXFP8Tensor is missing column-scaled FP8 data

(2) This happened after training completes and copying tensor to CPU:

1: [rank1]:   File "/usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/tensor/mxfp8_tensor.py", line 394, in _set_data                                                                           1: [rank1]:     new_device = tensor.device if tensor.is_cuda else self.device
1: [rank1]:                                   ^^^^^^^^^^^^^^
1: [rank1]: AttributeError: 'MXFP8TensorBase' object has no attribute 'is_cuda'

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
  [ ] Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Change A
• Change B

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[timmoon10]

This change makes the update_usage logic redundant, so we might as well remove it:

Suggested change

	if update_workspace and quantizer is not None and quantizer.columnwise_usage:
	tensor.update_usage(
	rowwise_usage=quantizer.rowwise_usage,
	columnwise_usage=quantizer.columnwise_usage,

That leads us to a more fundamental challenge: inference and validation have different data requirements, and we don't really expose a good way to distinguish between them. If you're doing MXFP8 inference, you don't want MXFP8 column-wise data since that's wasted memory. If you're loading a BF16 checkpoint, doing MXFP8 validation, and then doing MXFP8 training, you need to cast to MXFP8 row-wise and column-wise data when loading the checkpoint. Removing this update_usage logic has the effect of "however you've initialized the weights, you're stuck with those buffers forever". This is unfortunate for inference because the default behavior is to initialize weights for training, i.e. with both row-wise and column-wise data (#1827 does add an option to initialize for inference).

I think removing this logic does make sense eventually, especially after Mcore adds support for #1827, but for now I'm wary of breaking users doing inference. For now, is it possible to fix the bug by configuring distopt to avoid overlapping the param AG with validation?

[WanZzzzzz]

Thanks! Why do you think this bug has to do with overlapping the param AG with validation? I thought as long as we do training after validation, this error will occur.

[timmoon10]

Ah, you're right. I'll incorporate this logic in #1827.

[timmoon10]

I've included these fixes in #1827.